Quantification of Turnover Danger with xCounter

Abstract

Counterattacks in soccer are an important strategical component for goal scoring. Previous work in the literature has described their impact and has formulated descriptive advice on successful actions during a counterattack. In contrast, in this work, we propose the notion of expected counter, i.e., quantifying forward progress by the ball-winning team at the moment of the turnover. Therefore, we apply a previously proposed framework for understanding complex sequences in soccer. Using this framework, we perform a novel feature-specific assessment that yields (a) critical feature values, (b) relevant feature pitch zones, and (c) feature prediction capabilities. The insights from this assessment step allow for creating concrete guidelines for optimal behavior in and out of possession. Thus, we find that preparing horizontally spaced pass options facilitates an own counterattack in case of a ball win while moving as a compact unit prevents an opposing counterattack in case of a ball loss. As a final step, we generalize our results by creating a predictive XGBoost model that outperforms a location-based baseline but still shows room for improvement.

A Appendix

For the predictive XGBoost models, a range of different hyperparameters were examined. Therefore, training was done on the training set (70% of turnovers) and the performance of different hyperparameter configurations was compared on the test set (30% of turnovers). Due to the fact, that the focus of work does not lie on generating an automatic model, we avoid using a more sophisticated approach, e.g., using cross-validation or a separate validation set. A list of the evaluated hyperparameters is provided in Table 6. Admittedly, our search space is limited, yet we plan to expand it in future. An explanation of the different hyperparameter configurations of the best XGBoost models, encrypted as superscripts in Table 3, is given in Table 7.

Detailed lists of the best ranked features for both teams are presented in Tables 8 and 9.

Table 6. Hyperparameter search space used for optimizing the XGBoost models. The choice of search space was inspired by the optimal results documented by Bauer and Anzer [2] as experimentation with significantly different parameters did not offer promising results.

Table 7. Hyperparameter configurations of the superior XGBoost models (refer to Table 3 for evaluation results)

Table 8. Most predictive features for the ball-winning team. The Lastplayer subgroup was added to the feature set after inspection of the initial results.

Table 9. Most predictive features for the ball-losing team. The Lastplayer subgroup was added to the feature set after inspection of the initial results.